Contents
37 — Model organisms and genetic screens Part 8 — Methods

37 — Model organisms and genetic screens

Before this: Ch 09 · Ch 11 · Ch 14 · Ch 16 · Ch 25A — §9 assumes its gene-targeting and Cre-lox grammar, and §11 assumes its derivation · Ch 36 · Time: ~40 min

Sequencing tells you what is in the genome. It does not tell you what any of it is for. This chapter is about the experiments that build the sequence→function map, and about the handful of organisms in which those experiments are cheap enough to do exhaustively.

What you'll be able to do

The core idea

There are exactly two directions of inference between and , and every functional-genetics method is one of them.

starts from a phenotype and finds the gene. Break things at random, watch for the specific breakage you care about, then locate the lesion. You are fuzzing: you inject random , you have an oracle that says "this one is interesting", and the hard part afterwards is bisecting down to the responsible line. The enormous virtue is that you need no prior hypothesis. The enumerates the components of a process you do not understand, and it will hand you genes nobody would have nominated.

Reverse genetics starts from a gene and finds the phenotype. Delete or silence a chosen sequence and look for a difference. This is a point lookup with an unspecified return value. The lookup is now trivial — Ch 38 made it a reagent-ordering problem — and the difficulty has migrated entirely into the question what do I measure?

ForwardReverse
Starts fromPhenotypeGene
PerturbationRandom, unindexedTargeted
Hard stepFinding which mutation is causalChoosing and powering an assay
Prior hypothesis neededNone — this is the pointYes; you already picked the gene
Fails silently whenThe gene is redundant, essential, or pleiotropicThe assay does not touch the gene's function
YieldsComponents of a processConsequences of a component

Both directions bottom out in the same place: phenotyping. A screen is only as good as its oracle, and biology's oracles are weak, noisy, and low-dimensional.


1. Why any of this transfers: conservation, and where it stops

Studying yeast to learn about humans is only defensible because the core machinery is old and shared. Genes controlling the cell cycle were found in yeast by screening for temperature-sensitive division mutants; the human orthologues turned out to be so conserved that a human cDNA library can rescue a yeast mutant. The same is true across replication, , basic secretion, , and the main signalling cassettes. Roughly the last common ancestor's toolkit still runs, largely unmodified, in everything.

is deep where the machinery is ancient and shallow where evolution has been busy. It is worst exactly where translational medicine wants it most:

Conserved wellConserved poorly
Replication, repair, , translationImmune system detail, especially innate receptors
Cell cycle, cytoskeleton, membrane trafficDrug metabolism and xenobiotic response
Core signalling cassettes (RTK–Ras–MAPK, Wnt, Notch, Hedgehog)Regulatory wiring — when and where genes fire
Metabolic backboneBrain structure, lifespan, reproduction
Body-axis patterning genesRapidly evolving and lineage-expanded gene families (Ch 35)

A useful discipline: a is not a small human. It is an instrument with known distortion. Ask what the instrument is measuring before trusting the reading.

2. The organisms, and what each one actually buys

Counts are approximate and shift with each release; the human figure is pinned in verified-facts (GENCODE 50: 19,442 protein-coding genes).

SystemGeneration timeGenome~Protein-coding genesUniquely good forBlind spot
E. coli~20 min4.6 Mb~4,400Enormous population sizes; selections rather than ; biochemistryNo nucleus, no chromatin, no , no multicellularity
S. cerevisiae~90 min12 Mb~6,000Efficient makes precise gene replacement routine; switchable haploid/; complete deletion collectionNo development, no cell types, no tissues
C. elegans~3 days~100 Mb~20,000Invariant cell lineage — all 959 somatic nuclei traced from the zygote; transparent; 302-neuron connectome; selfing hermaphrodites homozygose for free; RNAi by feedingSmall, fast, weird: many pathways rewired; no adaptive immunity
D. melanogaster~10 days~180 Mb~14,000A century of ; balancer ; no meiotic recombination in males; polytene banding gave physical mapping decades earlyInsect-specific development and immunity
Zebrafish~3 months~1.4 Gb~26,000Vertebrate organogenesis watched live in a transparent externally-fertilised embryo; hundreds of embryos per clutchTeleost whole-genome duplication leaves paralogues that mask each other
Mouse~10 weeks~2.7 Gb~22,000Mammalian physiology, immunity, behaviour; -competent ES cells made targeted possible; inbred strains give genetic reproducibilitySlow, expensive; inbred strains are a single genetic background
A. thaliana~6–8 weeks~135 Mb~27,000Selfing, tiny, transformable by dipping flowers in Agrobacterium; indexed insertion collectionsPlant-specific everything
Human cell lineshours–days3.1 Gb19,442Human sequence, human proteins, scale for pooled screensUsually aneuploid and rearranged; no tissue context
Organoids / iPSCweeks3.1 Gb19,442Human with tissue architecture and cell-type diversityVariable, expensive, immature; no circulation or immune system

The pattern: you trade physiological relevance against throughput and genetic control, and each organism was adopted because it sat at a useful point on that curve. Nothing about the worm is intrinsically interesting; what is interesting is that you can screen a hundred thousand of them on nine agar plates and then trace the defect to a single named cell.

3. Forward screens I: mutagenesis

You need to break genes at a tunable rate and in a recoverable way.

MutagenLesion producedWhy you'd choose itCost
EMSAlkylates guanine; almost all G:C→A:T transitionsDense point ; produces an allelic series — nulls, partial-function hypomorphs, temperature-sensitivesThe lesion carries no tag; you must map it
ENUPoint mutations, mostly at A:T pairsHighest per-locus available in mouseSame mapping problem, at mouse cost
X-rays / γDouble-strand breaks → deletions, Clean nulls; deletions define intervals for mapping (Ch 18)Multi-gene lesions; rearrangements complicate breeding
Transposon / T-DNA insertionInsertion disrupting a geneThe mutation carries its own sequence tag — you recover the flanking DNA and you have the geneLower density; strong insertion-site bias; some genes never hit
librariesTargeted , genome-wide but indexedForward screening with reverse-genetic bookkeeping (Ch 38)Only targets what you designed guides for

The trade-off is legible: chemical mutagens give you high density and an allelic series but an expensive search afterwards; insertional mutagens give you the answer for free but sample the genome unevenly.

Dose is tuned, not maximised. The standard target is roughly one lethal-equivalent per mutagenised genome. Higher, and every mutant carries a dozen unrelated lesions you will spend months separating from the causal one; lower, and you more animals than you can afford.

4. Forward screens II: breeding schemes, and why recessives cost two extra generations

Most loss-of-function are , because most genes are haplosufficient — half the normal dose of product is enough. So the mutagenised individual, heterozygous for everything, shows nothing. An F1 finds only. To see a recessive you must make it homozygous, and that requires a breeding scheme.

MOUSE — three-generation ENU scheme for autosomal recessives

  P    ENU-treated male  ×  wild-type female
         |
  F1    +/m   (heterozygous, one per mutagenised gamete)   ← screen here = dominants only
         |  × wild-type female
  F2    1/2 of daughters are +/m                            ← carriers, invisible
         |  backcross F2 daughter × her F1 father
  F3    1/4 of progeny are m/m                              ← screen here = recessives

Three generations, ~30 weeks, and a colony that grows geometrically. This is the real reason mouse forward screens are rare and mouse reverse genetics is dominant.

The other organisms each cheat differently.

C. elegans self-fertilises. Every F1 hermaphrodite is a self-crossing heterozygote, so its brood is automatically 1/4 homozygous — the F2 screen is free, with no cross to set up. Picking F2s and scoring their F3 broods, the chance that a randomly chosen F2 is homozygous is 1/4, so to be 95% sure of catching at least one you need n with (3/4)ⁿ ≤ 0.05, i.e. n ≥ 11 clones per line.

Drosophila uses balancer , which are the cleverest piece of engineering in classical genetics. A balancer is a chromosome carrying multiple nested , a dominant visible marker, and a recessive lethal. The inversions mean any crossover with the homologue produces an inviable product, so recombinants are never recovered: the whole chromosome behaves as a single non-recombining allele. The recessive lethal means the balancer can never go homozygous. The marker means you can see who carries it. Put a mutagenised chromosome opposite a balancer and you get a stock that maintains itself as a permanent heterozygote forever, without selection and without the being lost or recombined away — a reference-counted immutable object, implemented in chromosomes in 1918.

5. Saturation: knowing when the screen is finished

If hit genes independently at mean rate λ per gene, the fraction of targetable genes hit at least once is 1 − e^(−λ). That is the textbook answer and it is optimistic, because mutability varies enormously between genes — a 300 bp gene with one critical residue is a far smaller target than a 5 kb gene.

The empirical test is better, and it is a species-richness problem — the same sampling logic as S3. Sort the mutants into complementation groups (§8) and look at the spectrum: how many groups are represented by one allele, two, three. A approaching saturation shows most genes hit repeatedly and almost no new groups appearing among the last mutants collected. The Chao-style estimator for unseen classes,

    unseen genes  ≈  f₁² / (2 f₂)

where f₁ is the number of single-allele groups and f₂ the number of two-allele groups, is exactly the same estimator you would use for unseen species — or unseen bugs in a fuzzing campaign. Many singletons and few doubletons means you are nowhere near done.

Saturation is a property of the screen, never of the genome. Four classes of gene are invisible to any screen no matter how deeply you push it: genes with redundant paralogues (the single mutant looks normal), genes required earlier than the process being scored (the animal dies before you can look), genes whose loss is pleiotropic enough that the mutant is filtered out for other reasons, and genes whose function the assay simply does not interrogate. "We saturated the screen" always means "for this , in this background, under these conditions".

6. Readouts, and the two screen designs that find pathways

The naive readout is a visible — the animal is the wrong shape, the colony is the wrong colour, the embryo lacks a structure. Three refinements do most of the real work.

Conditional . Essential genes cannot be recovered as nulls; the mutant is dead. A temperature-sensitive allele — typically a change that folds at 23 °C and misfolds at 37 °C — converts an essential gene into a switch. Shift the temperature, ask when in the process the requirement lies. The entire logic of the cell cycle was extracted this way, from yeast mutants that stop dividing at a specific, reproducible point when warmed. Note what a ts allele gives you that a never can: acute loss, with no time for compensation, and temporal resolution within a single cell cycle.

Suppressor . Start with a mutant, mutagenise it again, and screen for restored function. Suppressors are either intragenic (a second change in the same protein restoring the fold or an interaction surface — evidence about structure) or extragenic (a in a different gene that bypasses or compensates for the defect — evidence about the pathway).

/ modifier screens. Start from a weak allele that sits near the phenotypic threshold, then screen for second-site mutations that worsen it. This is the classic route to pathway components, and the reason is statistical: a sensitised background moves the assay onto the steep part of the dose–response curve, so a gene whose loss produces no detectable phenotype alone produces a large one here. Modifier screens select for functional connection rather than for having a phenotype at all, which is precisely the filter you want when the question is "what else is in this pathway?".

7. From mutant to gene

Classically: recombination mapping against marked strains to get a chromosomal interval (Ch 14), deletion mapping to narrow it, then transformation rescue — reintroduce a wild-type candidate and ask whether the disappears. Rescue is still the gold standard, because it is the only step that demonstrates sufficiency rather than correlation.

Modern practice collapses mapping and lesion-finding into one sequencing experiment. Bulk-segregant analysis is -frequency estimation on a selected pool:

graph LR A["mutant<br/>strain A background"] --> C["F1 hybrid"] B["polymorphic<br/>mapping strain B"] --> C C --> D["F2 segregants"] D --> E["select ~50–200 with<br/>the mutant phenotype"] E --> F["pool DNA,<br/>sequence to ~20×"] F --> G["strain-A allele frequency<br/>in windows along the genome"] G --> H["peak → interval;<br/>filter for mutagen-consistent<br/>coding changes"]

The estimator: unlinked markers sit at strain-A frequency 0.5, because half the genome came from each parent. At the causal locus every selected must carry the mutant allele, so the frequency goes to 1.0. At recombination fraction r from the locus the expected frequency is 1 − r. You are reading a genetic map directly off a -normalised allele-frequency track.

The critical design point, and the one that gets wrong: resolution is set by the number of selected meioses, not by sequencing depth. Per-site depth need only be enough that read-sampling noise, averaged across the many markers in a mapping window, falls below the pool-composition noise — which is why 20× over a 100-chromosome pool is ample. Depth cannot buy resolution the meioses do not contain: with N selected chromosomes the nearest recovered recombination breakpoint sits about 1/N Morgans away, so ~±1 cM is the floor for a 100-chromosome pool whether you sequence at 20× or 200× — and a pooled allele-frequency readout reaches that floor only if the pool is also large enough to estimate frequency that precisely (worked example below).

8. Complementation: same gene or different?

You have 62 mutants. How many genes is that? Cross two mutants and look at the heterozygous progeny (see Ch 11):

   m1/m1  ×  m2/m2   →   m1/m2

   progeny wild-type  → COMPLEMENTATION    → different genes
   progeny mutant     → NON-COMPLEMENTATION → same gene

The logic is that two broken copies of different genes each supply what the other lacks; two broken copies of the same gene supply nothing. The equivalence classes are complementation groups, and the count of groups is the count of genes.

Three ways it lies, all worth knowing: the test is only valid for recessive loss-of-function (a allele fails to complement everything); intragenic complementation can occur between alleles hitting different domains of a multimeric protein, producing spurious complementation within one gene; and unlinked non-complementation between genes whose products are dosage-sensitive partners produces spurious non-complementation between two genes. Confirm with sequence.

9. Reverse genetics

Targeted replacement. In yeast, is efficient enough that a product with 40 bp of on each end replaces an entire with a selectable marker. That single fact is why the systematic yeast deletion collection exists — one strain per gene, each carrying a unique 20 bp molecular barcode at the deleted locus, which turns out to matter enormously later (§10). In mammalian cells HR is rare, so the mouse route required a positive/negative selection cassette to enrich for correct targeting, and a cell type — embryonic stem cells — that could be manipulated in culture and still contribute to the of a chimeric animal.

Both paragraphs here are summaries. The mouse targeting route — vector design, the positive/negative selection logic that makes a rare event findable, ES cells, chimeras and germline transmission — and the conditional and inducible grammar built on Cre-lox are derived step by step in Ch 25A §§6–7. Read that first if the reasoning below is compressed to the point of assertion.

Conditional alleles. A straight has two failure modes: the animal dies before the tissue of interest exists, and the animal that survives has had its whole development to compensate. Site-specific recombination fixes both. Flank an essential exon with loxP sites, supply Cre recombinase from a tissue-specific , and the gene is deleted only where Cre is expressed. Fuse Cre to a modified oestrogen receptor and it stays out of the nucleus until tamoxifen is given, so you also choose when. The general principle — separate the allele from the trigger, then control the trigger in space and time — recurs everywhere in modern genetics.

Knockdown. RNAi (Ch 24) degrades or blocks the transcript. It is fast, dose-tunable, and works in organisms where targeting is hard — in C. elegans you can deliver it by feeding bacteria expressing double-stranded RNA. Its weakness is that specificity is set by a short seed match, so every reagent silences a partly-unknown set of other transcripts. Treat an RNAi as a multiple-hypothesis problem with strongly correlated errors: the standard controls are multiple independent reagents per gene and rescue with an RNAi-resistant transgene.

Morpholinos are antisense oligomers that block initiation or , and they were the zebrafish workhorse for a decade. Then systematic comparison of morphant against mutants in the same genes found that a large fraction did not reproduce. Two distinct explanations turned out to be true simultaneously, and disentangling them is a genuinely instructive episode: some morphant phenotypes were off-target artefacts, and some mutants had genetically compensated — degradation of the mutant transcript triggers transcriptional upregulation of related genes, so the mutant is less affected than an acute knockdown. Both effects are real. The lesson generalises far beyond fish: knockdown and knockout are not the same experiment, because a stable mutant has been selected for tolerance and an acute perturbation has not. Disagreement between them is information, not noise.

10. Pooled screens: differential abundance with a sequencing readout

This is where a modern becomes a counting problem, and where your statistics background pays for itself.

Build a library of perturbations — one construct per gene, or several — each carrying a unique barcode. Infect cells at low multiplicity so that most cells receive at most one construct; the integrated barcode is now a heritable label on that cell's lineage. Apply a selection. Sequence the barcode amplicon before and after. The of every gene is a change in the abundance of its barcode.

graph LR L["barcoded library<br/>~10⁵ constructs"] --> I["infect at low MOI<br/>≤1 construct per cell"] I --> T0["t₀ sample<br/>sequence barcodes"] I --> S["selection:<br/>growth / drug / FACS sort"] S --> T1["t₁ sample<br/>sequence barcodes"] T0 --> N["normalise counts"] T1 --> N N --> LFC["per-construct log₂ fold change"] LFC --> G["aggregate to gene-level statistic"]

Four things determine whether the result is real.

Counts are compositional. A construct's read fraction depends on every other construct in the pool, so a genuinely strong dropout inflates the apparent enrichment of everything else. The fix is the same family of used for (Ch 47) — size-factor or median-ratio scaling rather than raw proportions.

The error model lives at the construct level. Reagent efficiency varies wildly: of eight guides against one gene, two may be inert. Gene-level statistics therefore aggregate across constructs robustly — rank-based or trimmed rather than a plain mean — and the spread across constructs targeting the same gene is your best internal estimate of noise.

The null must be matched, not assumed. Non-targeting controls establish the sampling baseline. But in screens, cutting itself is toxic in proportion to copy number, so a gene in an amplified region drops out for reasons that have nothing to do with its function. The appropriate control cuts the genome a matched number of times at inert loci.

Cell , not read depth, sets the noise floor. Every passage is a multinomial resampling of the pool. If you carry only 50 cells per construct through a bottleneck, the count for a neutral construct wanders by tens of percent on alone, and the effect you are chasing is buried. Several hundred cells per construct at every step is the practical requirement, and the most common cause of an irreproducible screen is a purely sampling-statistical failure at a step nobody recorded.

11. Genetic interactions: epistasis and synthetic lethality

Single mutants tell you which genes matter. Double mutants tell you how genes relate.

Define the expectation under independence multiplicatively: if single mutants have relative Wₐ and W_b, the neutral expectation for the double is WₐW_b, and the interaction is

    ε  =  W_ab  −  W_a · W_b

    ε < 0   aggravating   (synthetic sick / synthetic lethal)  — parallel, buffering pathways
    ε > 0   alleviating   (buffering / masking)                — often the same complex or pathway
    ε ≈ 0   independent

Measured genome-wide in yeast across millions of double mutants, the matrix of ε values becomes a similarity graph over genes: cluster the rows and functional modules — complexes, pathways, organelles — fall out without anyone having annotated anything. Genes in the same complex share interaction profiles even when they interact weakly with each other.

is the therapeutically interesting corner. If losing gene A is survivable and losing gene B is survivable but losing both is not, then a tumour that has already lost A is selectively killable by a drug against B, while normal cells retaining A are not. Tumours defective in homologous-recombination repair — BRCA1- or BRCA2-mutant, for instance — are hypersensitive to inhibition of PARP, a protein central to handling single-strand DNA damage; inhibitors trap it on DNA and generate lesions that only can resolve. The general principle is the modifier relocated into oncology: the tumour's is a permanent sensitised background, and a genome-wide screen in that background enumerates candidate drug targets (Ch 56).

12. Phenotyping, background, and the meaning of "no phenotype"

Roughly a third of mouse gene are embryonic-lethal or subviable. Of the rest, a large fraction are reported as having no detectable — and that statement is almost never about the gene.

The genome offers ~20,000 columns. A well-funded phenotyping pipeline measures a few hundred parameters, under one diet, in one facility, in one genetic background, at one age. Absence of evidence is cheap to generate.

The specific reasons a real function goes undetected:

The defensible statement is never "gene X has no function". It is: "no phenotype was detected in assays A, B and C, in background D, at 80% power to detect an effect of size E." That sentence is longer and it is the only one that survives contact with a replication attempt.


Common misconceptions

What people thinkWhat's actually true
A with no means the gene does nothingIt means the assays used did not detect a difference. Redundancy, condition-specificity, compensation and low power all produce this result in genes with essential functions
are chosen because they resemble humansThey are chosen for throughput, genetic tractability and scorable phenotypes. Relevance is traded away deliberately and must be re-argued for each question
Knockdown and knockout answer the same questionThey differ systematically: stable mutants can compensate, acute knockdowns cannot. Disagreement between them is a real biological signal, not a technical failure
A saturated has found all the genes for a processIt has found all the genes findable by that assay, in that background. Redundant, essential and pleiotropic genes are structurally invisible
means physical interactionClassical epistasis orders genes in an information-flow sense. Two genes can be epistatic without their products ever touching, and interact physically with no epistasis
Complementation testing is a reliable way to count genesIt is reliable only for loss-of-function . , intragenic complementation, and unlinked non-commplementation all break it
Bigger screens need deeper sequencingResolution in bulk-segregant mapping is set by selected meioses, and noise in pooled screens by cells per construct. Sequencing depth is rarely the binding
made obsoleteIt made targeted perturbation cheap. Unbiased discovery still requires an assay that scores a phenotype, and that has not become easier

Worked example: one screen, from mutagen to pathway order

A C. elegans for animals that fail to respond to a chemical cue. Every number below is either standard or derived from the ones before it.

1. Mutagenesis and breeding. Soak L4 hermaphrodites in EMS. F1 animals are heterozygous for everything; self-progeny (F2) are 1/4 homozygous at any mutagenised locus. Clone F2s to individual plates and score F3 broods. To be 95% certain of recovering a homozygote from a given heterozygous line, clone n F2s with (3/4)ⁿ ≤ 0.05 → n ≥ ln(0.05)/ln(0.75) = 10.4, so 11 clones per line.

2. The haul, and saturation. The screen yields 62 mutants. Pairwise sort them into 27 groups, with the spectrum:

   alleles per group :  1    2    3    ≥4
   number of groups  :  9    7    5     6      →  9(1) + 7(2) + 5(3) + 6(≥4) = 62 alleles
                        ^    ^
                        f₁   f₂

Unseen genes ≈ f₁²/(2f₂) = 81/14 ≈ 6. Estimated total ≈ 33 genes; 27 recovered, so the screen is roughly 82% saturated. Nine singletons is the signal that it is not finished — another 20,000 haploid genomes is justified.

3. Mapping one mutant. Take m14, a singleton. Cross the mutant (Bristol background) to the polymorphic Hawaiian strain, self the F1, select 50 F2 animals with the mutant (= 100 selected ), pool their DNA and sequence to 20×. Compute Bristol in 1 Mb windows:

   region              expected Bristol frequency
   unlinked            0.50        (half the genome from each parent)
   10 cM from locus    1 − 0.10 = 0.90
    1 cM from locus    1 − 0.01 = 0.99
   at the locus        1.00        (selection guarantees it)

Binomial noise on 100 chromosomes is √(0.25/100) = 0.05, so the 0.90 window sits 8 SE above background — unmistakable. But separating 0.99 from 0.95 needs SE ≈ 0.01, requiring ~290 chromosomes. The peak is obvious with 50 animals; the fine interval needs ~150. Two limits are in play and they do not coincide: 1/N Morgans (~1 cM at N = 100) is the breakpoint-limited floor, but a pooled-frequency readout reaches it only with the ~290 chromosomes just computed, so with 100 chromosomes expect a peak spanning several cM. Even at the floor, ±1 cM ≈ ±300 kb in this genome — a 600 kb window holding roughly 120 genes (C. elegans averages ~1 gene per 5 kb), and more still if the interval falls in a gene-dense, low-recombination chromosome centre, where 1 cM buys well over a megabase. Filter that interval for EMS-consistent (G:C→A:T) changes altering coding sequence: typically one to three candidates. Confirm by transformation rescue.

4. Ordering two genes. The screen also produced a rare gain-of-function allele, b(gf), whose carriers respond constitutively — with no cue present. Gene a has a null, a(lf), whose carriers never respond. Build the double mutant and read the answer:

   PATHWAY MODEL (positive-acting chain):     cue → A → B → response

   a(lf)          : no response          (chain broken at A)
   b(gf)          : constitutive         (B fires without input)

   double a(lf); b(gf)
     ├─ observed CONSTITUTIVE  → B does not need A → B is DOWNSTREAM of A
     └─ observed NO RESPONSE   → B's output still requires A → B is UPSTREAM of A

Observed: constitutive. Therefore B acts downstream of A. Three conditions had to hold for this to mean anything: the alleles must be a true null and a true constitutive (a hypomorph gives an intermediate that orders nothing), the two mutants must have opposite, distinguishable phenotypes, and the pathway must be a linear chain of positive regulators. If either gene were a negative regulator the slogan "the epistatic mutant is downstream" inverts, and you must reason from the wiring rather than recite the rule.

Connections